2 * Copyright (c) 1982, 1986, 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * (c) UNIX System Laboratories, Inc.
5 * All or some portions of this file are derived from material licensed
6 * to the University of California by American Telephone and Telegraph
7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8 * the permission of UNIX System Laboratories, Inc.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 4. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
41 #include "opt_device_polling.h"
42 #include "opt_hwpmc_hooks.h"
44 #include "opt_watchdog.h"
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/callout.h>
50 #include <sys/kernel.h>
51 #include <sys/kthread.h>
54 #include <sys/mutex.h>
56 #include <sys/resource.h>
57 #include <sys/resourcevar.h>
58 #include <sys/sched.h>
59 #include <sys/signalvar.h>
60 #include <sys/sleepqueue.h>
64 #include <vm/vm_map.h>
65 #include <sys/sysctl.h>
67 #include <sys/interrupt.h>
68 #include <sys/limits.h>
69 #include <sys/timetc.h>
76 #include <sys/pmckern.h>
80 extern void hardclock_device_poll(void);
81 #endif /* DEVICE_POLLING */
83 static void initclocks(void *dummy);
84 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL);
86 /* Spin-lock protecting profiling statistics. */
87 static struct mtx time_lock;
90 sysctl_kern_cp_time(SYSCTL_HANDLER_ARGS)
93 long cp_time[CPUSTATES];
96 unsigned int cp_time32[CPUSTATES];
99 read_cpu_time(cp_time);
101 if (req->flags & SCTL_MASK32) {
103 return SYSCTL_OUT(req, 0, sizeof(cp_time32));
104 for (i = 0; i < CPUSTATES; i++)
105 cp_time32[i] = (unsigned int)cp_time[i];
106 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
111 return SYSCTL_OUT(req, 0, sizeof(cp_time));
112 error = SYSCTL_OUT(req, cp_time, sizeof(cp_time));
117 SYSCTL_PROC(_kern, OID_AUTO, cp_time, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
118 0,0, sysctl_kern_cp_time, "LU", "CPU time statistics");
120 static long empty[CPUSTATES];
123 sysctl_kern_cp_times(SYSCTL_HANDLER_ARGS)
130 unsigned int cp_time32[CPUSTATES];
136 if (req->flags & SCTL_MASK32)
137 return SYSCTL_OUT(req, 0, sizeof(cp_time32) * (mp_maxid + 1));
140 return SYSCTL_OUT(req, 0, sizeof(long) * CPUSTATES * (mp_maxid + 1));
142 for (error = 0, c = 0; error == 0 && c <= mp_maxid; c++) {
143 if (!CPU_ABSENT(c)) {
145 cp_time = pcpu->pc_cp_time;
150 if (req->flags & SCTL_MASK32) {
151 for (i = 0; i < CPUSTATES; i++)
152 cp_time32[i] = (unsigned int)cp_time[i];
153 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32));
156 error = SYSCTL_OUT(req, cp_time, sizeof(long) * CPUSTATES);
161 SYSCTL_PROC(_kern, OID_AUTO, cp_times, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,
162 0,0, sysctl_kern_cp_times, "LU", "per-CPU time statistics");
165 static const char *blessed[] = {
171 static int slptime_threshold = 1800;
172 static int blktime_threshold = 900;
173 static int sleepfreq = 3;
181 int blkticks, i, slpticks, slptype, tryl, tticks;
185 blkticks = blktime_threshold * hz;
186 slpticks = slptime_threshold * hz;
189 * Avoid to sleep on the sx_lock in order to avoid a possible
190 * priority inversion problem leading to starvation.
191 * If the lock can't be held after 100 tries, panic.
193 if (!sx_try_slock(&allproc_lock)) {
195 panic("%s: possible deadlock detected on allproc_lock\n",
198 pause("allproc", sleepfreq * hz);
202 FOREACH_PROC_IN_SYSTEM(p) {
204 if (p->p_state == PRS_NEW) {
208 FOREACH_THREAD_IN_PROC(p, td) {
211 * Once a thread is found in "interesting"
212 * state a possible ticks wrap-up needs to be
216 if (TD_ON_LOCK(td) && ticks < td->td_blktick) {
219 * The thread should be blocked on a
220 * turnstile, simply check if the
221 * turnstile channel is in good state.
223 MPASS(td->td_blocked != NULL);
225 tticks = ticks - td->td_blktick;
227 if (tticks > blkticks) {
230 * Accordingly with provided
231 * thresholds, this thread is
232 * stuck for too long on a
236 sx_sunlock(&allproc_lock);
237 panic("%s: possible deadlock detected for %p, blocked for %d ticks\n",
238 __func__, td, tticks);
240 } else if (TD_IS_SLEEPING(td) &&
242 ticks < td->td_blktick) {
245 * Check if the thread is sleeping on a
246 * lock, otherwise skip the check.
247 * Drop the thread lock in order to
248 * avoid a LOR with the sleepqueue
251 wchan = td->td_wchan;
252 tticks = ticks - td->td_slptick;
254 slptype = sleepq_type(wchan);
255 if ((slptype == SLEEPQ_SX ||
256 slptype == SLEEPQ_LK) &&
260 * Accordingly with provided
261 * thresholds, this thread is
262 * stuck for too long on a
264 * However, being on a
265 * sleepqueue, we might still
266 * check for the blessed
270 for (i = 0; blessed[i] != NULL;
272 if (!strcmp(blessed[i],
283 sx_sunlock(&allproc_lock);
284 panic("%s: possible deadlock detected for %p, blocked for %d ticks\n",
285 __func__, td, tticks);
292 sx_sunlock(&allproc_lock);
294 /* Sleep for sleepfreq seconds. */
295 pause("-", sleepfreq * hz);
299 static struct kthread_desc deadlkres_kd = {
302 (struct thread **)NULL
305 SYSINIT(deadlkres, SI_SUB_CLOCKS, SI_ORDER_ANY, kthread_start, &deadlkres_kd);
307 SYSCTL_NODE(_debug, OID_AUTO, deadlkres, CTLFLAG_RW, 0, "Deadlock resolver");
308 SYSCTL_INT(_debug_deadlkres, OID_AUTO, slptime_threshold, CTLFLAG_RW,
309 &slptime_threshold, 0,
310 "Number of seconds within is valid to sleep on a sleepqueue");
311 SYSCTL_INT(_debug_deadlkres, OID_AUTO, blktime_threshold, CTLFLAG_RW,
312 &blktime_threshold, 0,
313 "Number of seconds within is valid to block on a turnstile");
314 SYSCTL_INT(_debug_deadlkres, OID_AUTO, sleepfreq, CTLFLAG_RW, &sleepfreq, 0,
315 "Number of seconds between any deadlock resolver thread run");
316 #endif /* DEADLKRES */
319 read_cpu_time(long *cp_time)
324 /* Sum up global cp_time[]. */
325 bzero(cp_time, sizeof(long) * CPUSTATES);
328 for (j = 0; j < CPUSTATES; j++)
329 cp_time[j] += pc->pc_cp_time[j];
334 #include <sys/watchdog.h>
336 static int watchdog_ticks;
337 static int watchdog_enabled;
338 static void watchdog_fire(void);
339 static void watchdog_config(void *, u_int, int *);
340 #endif /* SW_WATCHDOG */
343 * Clock handling routines.
345 * This code is written to operate with two timers that run independently of
348 * The main timer, running hz times per second, is used to trigger interval
349 * timers, timeouts and rescheduling as needed.
351 * The second timer handles kernel and user profiling,
352 * and does resource use estimation. If the second timer is programmable,
353 * it is randomized to avoid aliasing between the two clocks. For example,
354 * the randomization prevents an adversary from always giving up the cpu
355 * just before its quantum expires. Otherwise, it would never accumulate
356 * cpu ticks. The mean frequency of the second timer is stathz.
358 * If no second timer exists, stathz will be zero; in this case we drive
359 * profiling and statistics off the main clock. This WILL NOT be accurate;
360 * do not do it unless absolutely necessary.
362 * The statistics clock may (or may not) be run at a higher rate while
363 * profiling. This profile clock runs at profhz. We require that profhz
364 * be an integral multiple of stathz.
366 * If the statistics clock is running fast, it must be divided by the ratio
367 * profhz/stathz for statistics. (For profiling, every tick counts.)
369 * Time-of-day is maintained using a "timecounter", which may or may
370 * not be related to the hardware generating the above mentioned
380 static DPCPU_DEFINE(int, pcputicks); /* Per-CPU version of ticks. */
381 static int global_hardclock_run = 0;
384 * Initialize clock frequencies and start both clocks running.
394 * Set divisors to 1 (normal case) and let the machine-specific
397 mtx_init(&time_lock, "time lock", NULL, MTX_DEF);
401 * Compute profhz/stathz, and fix profhz if needed.
403 i = stathz ? stathz : hz;
406 psratio = profhz / i;
408 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0);
413 * Each time the real-time timer fires, this function is called on all CPUs.
414 * Note that hardclock() calls hardclock_cpu() for the boot CPU, so only
415 * the other CPUs in the system need to call this function.
418 hardclock_cpu(int usermode)
420 struct pstats *pstats;
421 struct thread *td = curthread;
422 struct proc *p = td->td_proc;
426 * Run current process's virtual and profile time, as needed.
431 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
433 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
434 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
437 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
439 if (itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
440 flags |= TDF_PROFPEND | TDF_ASTPENDING;
445 td->td_flags |= flags;
449 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
450 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
456 * The real-time timer, interrupting hz times per second.
459 hardclock(int usermode, uintfptr_t pc)
462 atomic_add_int((volatile int *)&ticks, 1);
463 hardclock_cpu(usermode);
465 cpu_tick_calibration();
467 * If no separate statistics clock is available, run it from here.
469 * XXX: this only works for UP
472 profclock(usermode, pc);
475 #ifdef DEVICE_POLLING
476 hardclock_device_poll(); /* this is very short and quick */
477 #endif /* DEVICE_POLLING */
479 if (watchdog_enabled > 0 && --watchdog_ticks <= 0)
481 #endif /* SW_WATCHDOG */
485 hardclock_anycpu(int cnt, int usermode)
487 struct pstats *pstats;
488 struct thread *td = curthread;
489 struct proc *p = td->td_proc;
490 int *t = DPCPU_PTR(pcputicks);
491 int flags, global, newticks;
494 #endif /* SW_WATCHDOG */
497 * Update per-CPU and possibly global ticks values.
502 newticks = *t - global;
509 } while (!atomic_cmpset_int(&ticks, global, *t));
512 * Run current process's virtual and profile time, as needed.
517 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) {
519 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL],
521 flags |= TDF_ALRMPEND | TDF_ASTPENDING;
524 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) {
526 if (itimerdecr(&pstats->p_timer[ITIMER_PROF],
528 flags |= TDF_PROFPEND | TDF_ASTPENDING;
533 td->td_flags |= flags;
537 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
538 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
541 /* We are in charge to handle this tick duty. */
543 /* Dangerous and no need to call these things concurrently. */
544 if (atomic_cmpset_acq_int(&global_hardclock_run, 0, 1)) {
545 tc_ticktock(newticks);
546 #ifdef DEVICE_POLLING
547 /* This is very short and quick. */
548 hardclock_device_poll();
549 #endif /* DEVICE_POLLING */
550 atomic_store_rel_int(&global_hardclock_run, 0);
553 if (watchdog_enabled > 0) {
554 i = atomic_fetchadd_int(&watchdog_ticks, -newticks);
555 if (i > 0 && i <= newticks)
558 #endif /* SW_WATCHDOG */
560 if (curcpu == CPU_FIRST())
561 cpu_tick_calibration();
565 hardclock_sync(int cpu)
567 int *t = DPCPU_ID_PTR(cpu, pcputicks);
573 * Compute number of ticks in the specified amount of time.
579 register unsigned long ticks;
580 register long sec, usec;
583 * If the number of usecs in the whole seconds part of the time
584 * difference fits in a long, then the total number of usecs will
585 * fit in an unsigned long. Compute the total and convert it to
586 * ticks, rounding up and adding 1 to allow for the current tick
587 * to expire. Rounding also depends on unsigned long arithmetic
590 * Otherwise, if the number of ticks in the whole seconds part of
591 * the time difference fits in a long, then convert the parts to
592 * ticks separately and add, using similar rounding methods and
593 * overflow avoidance. This method would work in the previous
594 * case but it is slightly slower and assumes that hz is integral.
596 * Otherwise, round the time difference down to the maximum
597 * representable value.
599 * If ints have 32 bits, then the maximum value for any timeout in
600 * 10ms ticks is 248 days.
614 printf("tvotohz: negative time difference %ld sec %ld usec\n",
618 } else if (sec <= LONG_MAX / 1000000)
619 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
621 else if (sec <= LONG_MAX / hz)
623 + ((unsigned long)usec + (tick - 1)) / tick + 1;
632 * Start profiling on a process.
634 * Kernel profiling passes proc0 which never exits and hence
635 * keeps the profile clock running constantly.
639 register struct proc *p;
642 PROC_LOCK_ASSERT(p, MA_OWNED);
643 if (p->p_flag & P_STOPPROF)
645 if ((p->p_flag & P_PROFIL) == 0) {
646 p->p_flag |= P_PROFIL;
647 mtx_lock(&time_lock);
648 if (++profprocs == 1)
649 cpu_startprofclock();
650 mtx_unlock(&time_lock);
655 * Stop profiling on a process.
659 register struct proc *p;
662 PROC_LOCK_ASSERT(p, MA_OWNED);
663 if (p->p_flag & P_PROFIL) {
664 if (p->p_profthreads != 0) {
665 p->p_flag |= P_STOPPROF;
666 while (p->p_profthreads != 0)
667 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE,
669 p->p_flag &= ~P_STOPPROF;
671 if ((p->p_flag & P_PROFIL) == 0)
673 p->p_flag &= ~P_PROFIL;
674 mtx_lock(&time_lock);
675 if (--profprocs == 0)
677 mtx_unlock(&time_lock);
682 * Statistics clock. Updates rusage information and calls the scheduler
683 * to adjust priorities of the active thread.
685 * This should be called by all active processors.
688 statclock(int usermode)
700 cp_time = (long *)PCPU_PTR(cp_time);
703 * Charge the time as appropriate.
706 if (p->p_nice > NZERO)
712 * Came from kernel mode, so we were:
713 * - handling an interrupt,
714 * - doing syscall or trap work on behalf of the current
716 * - spinning in the idle loop.
717 * Whichever it is, charge the time as appropriate.
718 * Note that we charge interrupts to the current process,
719 * regardless of whether they are ``for'' that process,
720 * so that we know how much of its real time was spent
721 * in ``non-process'' (i.e., interrupt) work.
723 if ((td->td_pflags & TDP_ITHREAD) ||
724 td->td_intr_nesting_level >= 2) {
730 if (!TD_IS_IDLETHREAD(td))
737 /* Update resource usage integrals and maximums. */
738 MPASS(p->p_vmspace != NULL);
741 ru->ru_ixrss += pgtok(vm->vm_tsize);
742 ru->ru_idrss += pgtok(vm->vm_dsize);
743 ru->ru_isrss += pgtok(vm->vm_ssize);
744 rss = pgtok(vmspace_resident_count(vm));
745 if (ru->ru_maxrss < rss)
747 KTR_POINT2(KTR_SCHED, "thread", sched_tdname(td), "statclock",
748 "prio:%d", td->td_priority, "stathz:%d", (stathz)?stathz:hz);
749 thread_lock_flags(td, MTX_QUIET);
755 profclock(int usermode, uintfptr_t pc)
766 * Came from user mode; CPU was in user state.
767 * If this process is being profiled, record the tick.
768 * if there is no related user location yet, don't
769 * bother trying to count it.
771 if (td->td_proc->p_flag & P_PROFIL)
772 addupc_intr(td, pc, 1);
777 * Kernel statistics are just like addupc_intr, only easier.
780 if (g->state == GMON_PROF_ON && pc >= g->lowpc) {
782 if (i < g->textsize) {
791 * Return information about system clocks.
794 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS)
796 struct clockinfo clkinfo;
798 * Construct clockinfo structure.
800 bzero(&clkinfo, sizeof(clkinfo));
803 clkinfo.profhz = profhz;
804 clkinfo.stathz = stathz ? stathz : hz;
805 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
808 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate,
809 CTLTYPE_STRUCT|CTLFLAG_RD|CTLFLAG_MPSAFE,
810 0, 0, sysctl_kern_clockrate, "S,clockinfo",
811 "Rate and period of various kernel clocks");
816 watchdog_config(void *unused __unused, u_int cmd, int *error)
820 u = cmd & WD_INTERVAL;
821 if (u >= WD_TO_1SEC) {
822 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz;
823 watchdog_enabled = 1;
826 watchdog_enabled = 0;
831 * Handle a watchdog timeout by dumping interrupt information and
832 * then either dropping to DDB or panicking.
845 nintr = sintrcnt / sizeof(u_long);
847 printf("interrupt total\n");
848 while (--nintr >= 0) {
850 printf("%-12s %20lu\n", curname, *curintr);
851 curname += strlen(curname) + 1;
852 inttotal += *curintr++;
854 printf("Total %20ju\n", (uintmax_t)inttotal);
856 #if defined(KDB) && !defined(KDB_UNATTENDED)
858 kdb_enter(KDB_WHY_WATCHDOG, "watchdog timeout");
860 panic("watchdog timeout");
864 #endif /* SW_WATCHDOG */